reflow

Features

Everything Reflow does serves one outcome: an AI-authored change you can merge because it has been exercised in a real browser.

mcp

Agent-native by design

The MCP server is the primary interface. Tool descriptions alone teach a fresh agent the whole workflow — author a flow, trigger a run, read the result, commit the fix. The dashboard is the human window onto the same data.

runner

Real browsers, real proof

Every run executes in a real Chromium, Firefox, or WebKit browser on isolated cloud runners. Pass/fail comes with screenshots, traces, and visual diffs — evidence a reviewer can see, not a green dot to take on faith.

rfl

Flows that survive redesigns

Steps are plain language. Agents maintain RFL — a deterministic, human-readable step language that maps 1:1 onto Playwright — so green runs execute with no model in the loop. When the page changes, the agent regenerates just the affected step; the diff lands in your PR.

byo-llm

Bring your own model

Register your Anthropic or OpenAI key; runners decrypt it only for the duration of a run. You pay your provider directly — Reflow never marks up tokens and never trains on your data.

ci

PR checks that mean something

The GitHub Action runs your flows against each PR preview URL and posts a check with a per-flow table: status, duration, healed steps, artifacts. New push cancels the stale run automatically.

registry

Git stays the source of truth

Flows are markdown files in your repo. Reflow caches them content-addressed — like an artifact registry — but never holds a write token unless you install the GitHub App for one-click commits. Promotion is just your merge.

one-click

Approve & commit, in the PR

Healed flows and changed visual baselines arrive as a reviewable delta inside the PR comment. One button approves and commits them to the branch — opt-in via the GitHub App, never silent.

visual diff

Screenshot & component deltas

Every run diffs screenshots against the main-branch baseline — full page and per component. The check shows what moved, by how much, and whether the flow’s intent still holds.

agent inbox

An inbox for every run

Each run gets a disposable address at inbox.reflow.io. Signup verification, magic links, and OTP codes work without test-mode backdoors — the agent reads the mail and carries on.

context

Context packs

Drop your design style guide, domain glossary, and known quirks into .reflow/context/. The agent uses them to judge what "correct" looks like — in your product’s terms, not generic heuristics.

The loop, end to end

  1. 01 — Agent (or you) writes .reflow/flows/*.md next to the code it verifies.
  2. 02 — PR opens. The Action submits flows + preview URL to Reflow with an idempotency key per commit.
  3. 03 — Runners execute each flow. Scripted steps run deterministically; deviations, declared invariants, and final-state validation invoke your model.
  4. 04 — Check lands: green with screenshots, or red with a proposed rewrite your agent fetches over MCP and commits to the branch.